Projection-Based Methods in Optimization

Transcription

1 Projection-Based Methods in Optimization Charles Byrne (Charles Department of Mathematical Sciences University of Massachusetts Lowell Lowell, MA 01854, USA June 3, 2013

2 Talk Available on Web Site This slide presentation and accompanying article are available on my web site, ; click on Talks.

3 Using Prior Knowledge Figure : Minimum Two-Norm and Minimum Weighted Two-Norm Reconstruction.

4 Fourier Transform as Data Figure : Far-field Measurements. The distance from x to P is approximately D x cos θ.

5 Far-field Measurements Each point x in [ L, L] sends out the signal f (x) exp(iωt). The known frequency ω is the same for each x. We must estimate the function f (x), for x L. What a point P in the far-field receives from each x is approximately f (x) exp(ixω cos(θ)), so our measurement at P provides an approximation of L L f (x) exp(ixω cos(θ))dx. For c the propagation speed and λ the wavelength, if we have cos(θ) = nπc ωl = n λ 2L, then we have the nth Fourier coefficient of the function f (x).

6 The Limited-Data Problem We can have only if cos(θ) = nπc ωl = n λ 2L n 2L λ, so we can measure only a limited number of the Fourier coefficients of f (x). Note that the upper bound on n is the length of the interval [ L, L] in units of wavelength.

7 Can We Get More Data? Clearly, we can take our measurements at any point P on the circle, not just at those satisfying the equation cos(θ) = nπc ωl = n λ 2L. These measurements provide additional information about f (x), but won t be additional Fourier coefficients for the Fourier series of f (x) on [ L, L]. How can we use these additional measurements to improve our estimate of f (x)?

8 Over-Sampled Data Suppose that we over-sample; let us take measurements at points P such that the associated angle θ satisfies cos(θ) = nπc ωkl, where K > 1 is a positive integer, instead of cos(θ) = nπc ωl. Now we have Fourier coefficients for the function g(x) that is f (x) for x L, and is zero on the remainder of the interval [ KL, KL].

9 Using Support Information Given our original limited data, we can calculate the orthogonal projection of the zero function onto the subset of all functions on [ L, L] consistent with this data; this is the minimum-norm estimate of f (x), also called the discrete Fourier transform (DFT) estimate, shown in the second graph below. If we use the over-sampled data as Fourier coefficients for g(x) on the interval [ KL, KL], we find that we haven t improved our estimate of f (x) for x L. Instead, we calculate the orthogonal projection of the zero function onto the subset of all functions on [ L, L] that are consistent with the over-sampled data. This is sometimes called the modified DFT (MDFT) estimate. The top graph below shows the MDFT for a case of K = 30.

10 Two Minimum-Norm Estimates Figure : The non-iterative band-limited extrapolation method (MDFT) (top) and the DFT (bottom); 30 times over-sampled.

11 Over-Sampled Fourier-Transform Data For the simulation in the figure above, f (x) = 0 for x > L. The top graph is the minimum-norm estimator, with respect to the Hilbert space L 2 ( L, L), called the modified DFT (MDFT); the bottom graph is the DFT, the minimum-norm estimator with respect to the Hilbert space L 2 ( 30L, 30L), shown only for [ L, L]. The MDFT is a non-iterative variant of Gerchberg-Papoulis band-limited extrapolation.

12 Using Orthogonal Projection In both of the examples above we see minimum two-norm solutions consistent with the data. These reconstructions involve the orthogonal projection of the zero vector onto the set of solutions consistent with the data. The improvements illustrate the advantage gained by the selection of an appropriate ambient space within which to perform the projection. The constraints of data consistency define a subset onto which we project, and the distance to zero is the function being minimized, subject to the constraints. This leads to the more general problem of optimizing a function, subject to constraints.

13 The Basic Problem The basic problem we consider here is to minimize a real-valued function f : X R, over a subset C X, where X is an arbitrary set. With ι C (x) = 0, for x C, and +, otherwise, we can rewrite the problem as minimizing f (x) + ι C (x), over all x X.

14 The Limited-Data Problem We want to reconstruct or estimate a function of several continuous variables. We have limited noisy measured data and some prior information that serve to constrain that function. We discretize the function, so that the object is described as an unknown vector x in R J.

15 Constraints We want our reconstruction to be at least approximately consistent with the measured data. We may know that the entries of x are non-negative. When x is a vectorized image, we may have prior knowledge of its general appearance (a head slice, for example). If x is an extrapolated band-limited time series, we may have prior knowledge of the extent of the band.

16 Representing Constraints Prior knowledge of general properties of x can be incorporated through the choice of the ambient space. Other constraints, such as the measured data, tell us that x lies in a subset C of the ambient space.

17 An Example: Linear Functional Data Suppose that the measured data vector b is linear in x, with Ax = b under-determined. The minimum two-norm solution minimizes J x j 2, j=1 subject to Ax = b. Let the vector p with entries p j > 0 be a prior estimate of the magnitudes x j. The minimum weighted two-norm solution minimizes subject to Ax = b. J x j 2 /p j, j=1

18 Using Optimization When the constraint set C is relatively small, any member of C may provide an adequate solution to the problem. More likely, C is relatively large, and we may choose to determine x by minimizing a cost function over the set C.

19 An Example- Systems of Linear Equations Let the data b pertaining to x be linear, so that Ax = b for some matrix A. 1. When there are infinitely many solutions, we may select x to minimize a norm or other suitable distance measure; 2. When there are no solutions, we may select a least-squares or other approximate solution; 3. When the data b is noisy, we may select a regularized solution.

20 Barrier Functions: An Example The problem is to minimize the function f (x) = f (x 1, x 2 ) = x1 2 + x 2 2, subject to x 1 + x 2 1. For each positive integer k, the vector x k with entries x k 1 = x k 2 = k minimizes the function B k (x) = x1 2 + x k log(x 1 + x 2 1) = f (x) + 1 k b(x). Notice that x1 k + x 2 k > 1, so each x k satisfies the constraint. As k +, x k converges to ( 1 2, 1 2 ), which solves the original problem.

21 Penalty Functions: An Example The problem is to minimize the function Our penalty function is f (x) = (x + 1) 2, subject to x 0. p(x) = x 2, for x 0, p(x) = 0, for x > 0. At the kth step we minimize f (x) + kp(x) to get x k = 1 k+1, which converges to the right answer, x = 0, as k. The limit x satisfies the constraint, but the x k do not; this is an exterior-point method.

22 Constrained Optimization The Problem: to minimize a function f : X (, ], over a non-empty subset C of X, where X is an arbitrary set. Auxiliary-Function Methods: At the k th step of an auxiliary-function (AF) algorithm we minimize a function G k (x) = f (x) + g k (x) over x C to get x k. Auxiliary-function algorithms are closely related to sequential unconstrained minimization (SUM) methods. Several SUM methods, such as barrier-function and penalty-function methods, can be reformulated as AF methods.

23 Properties of Auxiliary Functions Auxiliary functions g k have the properties 1. g k (x) 0, for all x C; 2. g k (x k 1 ) = 0. Therefore, we have the following theorem. Theorem Let {x k } be generated by an AF algorithm. Then the sequence {f (x k )} is non-increasing. Proof: We have f (x k 1 ) = G k (x k 1 ) G k (x k ) f (x k ).

24 GOALS: The vector x k minimizes G k (x) = f (x) + g k (x), over x C. 1. We would like for the sequence {x k } to converge to some x C that solves the problem. This requires a topology on X. 2. Failing that, we would like the sequence {f (x k )} to converge to d = inf{f (x) x C}. 3. At the very least, we want the sequence {f (x k )} to be non-increasing.

25 SUMMA An AF algorithm in said to be in the SUMMA class if for all x C. Theorem G k (x) G k (x k ) g k+1 (x) 0, For algorithms in the SUMMA class, the sequence {f (x k )} is non-increasing, and we have {f (x k )} d = inf{f (x) x C}.

26 SUMMA Proof: If f (x k ) d > f (z) d for all k, then g k (z) g k+1 (z) g k (z) (G k (z) G k (x k )) = f (x k ) f (z) + g k (x k ) d f (z) > 0. But the decreasing non-negative sequence {g k (z)} cannot have its successive differences bounded away from zero.

27 Examples of SUMMA A number of well known algorithms either are in the SUMMA class, or can be reformulated to be in the SUMMA class, including 1. Barrier-function methods; 2. Penalty-function methods; 3. Forward-backward splitting (the CQ algorithm, projected Landweber, projected gradient descent); 4. Alternating minimization with the 5-point property (simultaneous MART); 5. Certain cases of the EM algorithm; 6. Proximal minimization with Bregman distances; 7. Statistical majorization minimization.

28 Barrier-Function Methods A function b : C [0, + ] is a barrier function for C, that is, b is finite on C, b(x) = +, for x not in C, and, for topologized X, has the property that b(x) + as x approaches the boundary of C. At the kth step of the iteration we minimize f (x) + 1 k b(x) to get x k. Equivalently, we minimize with G k (x) = f (x) + g k (x), g k (x) = [(k 1)f (x) + b(x)] [(k 1)f (x k 1 ) + b(x k 1 )]. Then G k (x) G k (x k ) = g k+1 (x).

29 Penalty-Function Methods We select a non-negative function p : X R with the property that p(x) = 0 if and only if x is in C and then, for each positive integer k, we minimize f (x) + kp(x), or, equivalently, p(x) + 1 k f (x), to get x k. Most, but not all, of what we need concerning penalty-function methods can then be obtained from the discussion of barrier-function methods.

30 Bregman Distances Let f : Z R J R be convex on its domain Z, and differentiable on the interior U of Z. For x Z and y U, the Bregman distance from y to x is D f (x, y) = f (x) f (y) f (y), x y. Then, because of the convexity of f, D f (x, y) 0 for all x and y. The Euclidean distance is the Bregman distance for f (x) = 1 2 x 2 2.

31 Proximal Minimization Algorithms (PMA) Let f : C R J R be convex on C and differentiable on the interior U of C. At the kth step of a proximal minimization (prox min) algorithm (PMA), we minimize the function G k (x) = f (x) + D h (x, x k 1 ), to get x k. The function g k (x) = D h (x, x k 1 ) is the Bregman distance associated with the Bregman function h. We assume that each x k lies in U, whose closure is the set C. We have G k (x) G k (x k ) = D f (x, x k ) + D h (x, x k ) D h (x, x k ) = g k+1 (x).

32 Another Job for the PMA Suppose that f (x) = 1 2 Ax b 2 2, with AA T invertible. If the PMA sequence {x k } converges to some x and h(x 0 ) is in the range of A T then x minimizes h(x) over all x with Ax = b.

33 Obstacles in the PMA To obtain x k in the PMA we must solve the equation f (x k ) + h(x k ) = h(x k 1 ). This is usually not easy. However, we can modify the PMA to overcome this obstacle. This modified PMA is an interior-point algorithm that we have called the IPA.

34 The IPA We still get x k by minimizing G k (x) = f (x) + D h (x, x k 1 ). With a(x) = h(x) + f (x), we find that x k solves a(x k ) = a(x k 1 ) f (x k 1 ). Therefore, we look for a(x) so that 1. h(x) = a(x) f (x) is convex; 2. obtaining x k from a(x k ) is easy now.

35 The IPA and Constraints In the PMA, h can be chosen to force x k to be in U, the interior of the domain of h. In the IPA it is a(x) that we select, and incorporating the constraints within a(x) may not be easy.

36 The Projected Landweber Algorithm as IPA We want to minimize f (x) = 1 2 Ax b 2 2, over x C. We select γ so that 0 < γ < 1 ρ(a T A) and We have and we minimize a(x) = 1 2γ x 2 2. D f (x, y) = 1 2 Ax Ay 2 2, G k (x) = f (x) + 1 2γ x x k Ax Ax k over x in C to get x k = P C (x k 1 γa T (Ax k 1 b)).

37 An Example of the IPA: Projected Gradient Descent Let f be convex and differentiable on R J, with f L-Lipschitz, and 0 < γ 1 L. Let C RJ be closed, nonempty, and convex. Let a(x) = 1 2γ x 2 2. At the kth step we minimize G k (x) = f (x) + 1 2γ x x k D f (x, x k 1 ), over x C, obtaining x k = P C (x k 1 γ f (x k 1 )); P C is the orthogonal projection onto C.

38 Projected Gradient Descent as SUMMA The auxiliary function g k (x) can be written as g k (x) = 1 2γ x x k D f (x, x k 1 ) = D h (x, x k 1 ), where h(x) is the convex differentiable function Then h(x) = a(x) f (x) = 1 2γ x 2 2 f (x). G k (x) G k (x k ) = D a (x, x k ) D h (x, x k ) = g k+1 (x).

39 Majorization Minimization or Optimization Transfer Majorization minimization or optimization transfer is used in statistical optimization. For each fixed y, let g(x y) f (x), for all x, and g(y y) = f (y). At the kth step we minimize g(x x k 1 ) to get x k. This statistical method is equivalent to minimizing where f (x) + D(x, x k 1 ) = f (x) + g k (x), g k (x) = D(x, x k 1 ), for some distance measure D(x, z) 0, with D(z, z) = 0.

40 The Method of Auslander and Teboulle The method of Auslander and Teboulle is a particular example of an MM method that is not in SUMMA. At the kth step of their method minimize G k (x) = f (x) + D(x, x k 1 ) to get x k. They assume that D has an associated induced proximal distance H(a, b) 0, finite for a and b in C, with H(a, a) = 0 and 1 d(b, a), c b H(c, a) H(c, b), for all c in C. Here 1 D(x, y) denotes the x gradient. We do have {f (x k )} d. If D = D h, then H = D h also; Bregman distances are self-proximal.

41 Projected Gradient Descent (PGD) Revisited At the kth step of the PGD we minimize G k (x) = f (x) + 1 2γ x x k D f (x, x k 1 ), over x C. Equivalently, we minimize the function ι C (x) + f (x) + 1 2γ x x k D f (x, x k 1 ), over all x in R J, with ι C (x) = 0 for x C and ι C (x) = + otherwise. Now the objective function is a sum of two functions, one non-differentiable. The forward-backward splitting method then applies.

42 Moreau s Proximity Operators Let f : R J R be convex. For each z R J the function m f (z) := min x {f (x) x z 2 2 } is minimized by a unique x = prox f (z). Moreau s proximity operator prox f extends the notion of orthogonal projection onto a closed convex set: if f (x) = ι C (x) then prox f (x) = P C (x). Also x = prox f (z) iff z x f (x) iff x = J f (z), where J f (z) is the resolvent of the set-valued operator f.

43 Forward-Backward Splitting Let f : R J R be convex, with f = f 1 + f 2, both convex, f 2 differentiable, and f 2 L-Lipschitz continuous. The iterative step of the FBS algorithm is ) x k = prox γf1 (x k 1 γ f 2 (x k 1 ), which can be obtained by minimizing G k (x) = f (x) + 1 2γ x x k D f 2 (x, x k 1 ). Convergence of the sequence {x k } to a solution can be established, if γ is chosen to lie within the interval (0, 1/L].

44 Projected Gradient Descent as FBS To put the projected gradient descent method into the framework of the forward-backward splitting we let f 1 (x) = ι C (x), the indicator function of the set C, which is zero for x C and + otherwise. Then we minimize the function over all x R J. f 1 (x) + f 2 (x) = ι C (x) + f (x)

45 The CQ Algorithm Let A be a real I by J matrix, C R J, and Q R I, both closed convex sets. The split feasibility problem (SFP) is to find x in C such that Ax is in Q. The function f 2 (x) = 1 2 P QAx Ax 2 2 is convex and differentiable, f 2 is ρ(a T A)-Lipschitz, and f 2 (x) = A T (I P Q )Ax. We want to minimize the function f (x) = ι C (x) + f 2 (x), over all x. The FBS algorithm gives the iterative step for the CQ algorithm; with 0 < γ 1/L, x k = P C (x k 1 γa T (I P Q )Ax k 1).

46 Alternating Minimization Suppose that P and Q are arbitrary non-empty sets and the function Θ(p, q) satisfies < Θ(p, q) +, for each p P and q Q. The general AM method proceeds in two steps: we begin with some q 0, and, having found q k 1, we 1. minimize Θ(p, q k 1 ) over p P to get p = p k, and then 2. minimize Θ(p k, q) over q Q to get q = q k. The 5-point property of Csiszár and Tusnády is Θ(p, q) + Θ(p, q k 1 ) Θ(p, q k ) + Θ(p k, q k 1 ).

47 AM as SUMMA When the 5-point property holds for AM, the sequence {Θ(p k, q k )} converges to d = inf Θ(p, q). p,q For each p P, define q(p) to be some member of Q for which Θ(p, q(p)) Θ(p, q), for all q Q; then q(p k ) = q k. Define f (p) = Θ(p, q(p)). At the kth step of AM we minimize where G k (p) = Θ(p, q k 1 ) = f (p) + g k (p), g k (p) = Θ(p, q k 1 ) Θ(p, q(p)). The 5-point property is then the SUMMA condition G k (p) G k (p k ) g k+1 (p) 0.

48 Kullback-Leibler Distances For a > 0 and b > 0, the Kullback-Leibler distance from a to b is KL(a, b) = a log a a log b + b a, with KL(0, b) = b, and KL(a, 0) = +. The KL distance is extended to non-negative vectors entry-wise. Let y R I be positive, and P an I by J matrix with non-negative entries. The simultaneous MART (SMART) algorithm minimizes KL(Px, y), and the EMML algorithm minimizes KL(y, Px), both over all non-negative x R J. Both algorithms can be viewed as particular cases of alternating minimization.

49 The General EM Algorithm The general EM algorithm is essentially non-stochastic. Let Z be an arbitrary set. We want to maximize L(z) = b(x, z)dx, over z Z, where b(x, z) : R J Z [0, + ]. Given z k 1, we take f (z) = L(z) and minimize G k (z) = f (z) + KL(b(x, z k 1 ), b(x, z))dx to get z k. Since g k (z) = KL(b(x, z k 1 ), b(x, z))dx 0, for all z Z and g k (z k 1 ) = 0, we have an AF method.

50 KL Projections For a fixed x, minimizing the distance KL(z, x) over z in the hyperplane H i = {z (Pz) i = y i } generally cannot be done in closed form. However, assuming that I i=1 P ij = 1, the weighted KL projections z i = T i (x) onto hyperplanes H i obtained by minimizing J P ij KL(z j, x j ) j=1 over z in H i, are given in closed form by y i zj i = T i (x) j = x j, for j = 1,..., J. (Px) i

51 SMART and EMML Having found x k, the next vector in the SMART sequence is x k+1 j = I (T i (x k ) j ) P ij, i=1 so x k+1 is a weighted geometric mean of the T i (x k ). For the EMML algorithm the next vector in the EMML sequence is I x k+1 j = P ij T i (x k ) j, i=1 so x k+1 is a weighted arithmetic mean of the T i (x k ).

52 The MART The MART algorithm has the iterative step x k+1 j where i = k(mod I) + 1 and = x k j (y i /(Px k ) i ) P ij m 1 i, m i = max{p ij j = 1, 2,..., J}. We can express the MART in terms of the weighted KL projections T i (x k ); x k+1 j = (x k j ) 1 P ij m 1 i (T i (x k ) j ) P ij m 1 i. We see then that the iterative step of the MART is a relaxed weighted KL projection onto H i, and a weighted geometric mean of the current x k j and T i (x k ) j.

53 The EMART The iterative step of the EMART algorithm is x k+1 j = (1 P ij m 1 i )xj k + P ij m 1 i T i (x k ) j. We can express the EMART in terms of the weighted KL projections T i (x k ); x k+1 j = (1 P ij m 1 i )xj k + P ij m 1 i T i (x k ) j. We see then that the iterative step of the EMART is a relaxed weighted KL projection onto H i, and a weighted arithmetic mean of the current x k j and T i (x k ) j.

54 MART and EMART Compared When there are non-negative solutions of the system y = Px, the MART sequence {x k } converges to the solution x that minimizes KL(x, x 0 ). The EMART sequence {x k } also converges to a non-negative solution, but nothing further is known about this solution. One advantage that the EMART has over the MART is the substitution of multiplication for exponentiation.

55 SMART as SUMMA At the kth step of the SMART we minimize the function G k (x) = KL(Px, y) + KL(x, x k 1 ) KL(Px, Px k 1 ) to get x k with entries x k j = x k 1 j exp ( I i=1 ) P ij log(y i /(Px k 1 ) i ). We assume that P and x have been rescaled so that I i=1 P ij = 1, for all j. Then g k (x) = KL(x, x k 1 ) KL(Px, Px k 1 )) 0. and G k (x) G k (x k ) = KL(x, x k ) g k+1 (x) 0.

56 SMART as IPA With f (x) = KL(Px, y), the associated Bregman distance is D f (x, z) = KL(Px, Pz). With we have a(x) = J x j log(x j ) x j, j=1 D a (x, z) = KL(x, z) KL(Px, Pz) = D f (x, z). Therefore, h(x) = a(x) f (x) is convex.

57 Convergence of the FBS For each k = 1, 2,... let where G k (x) = f (x) + 1 2γ x x k D f 2 (x, x k 1 ), D f2 (x, x k 1 ) = f 2 (x) f 2 (x k 1 ) f 2 (x k 1 ), x x k 1. Here D f2 (x, y) is the Bregman distance formed from the function f 2. The auxiliary function can be rewritten as g k (x) = 1 2γ x x k D f 2 (x, x k 1 ) g k (x) = D h (x, x k 1 ), where h(x) = 1 2γ x 2 2 f 2(x).

58 Proof (p.2): Therefore, g k (x) 0 whenever h(x) is a convex function. We know that h(x) is convex if and only if h(x) h(y), x y 0, for all x and y. This is equivalent to 1 γ x y 2 2 f 2(x) f 2 (y), x y 0. Since f 2 is L-Lipschitz, the inequality holds for 0 < γ 1/L.

59 Proof (p.3): Lemma The x k that minimizes G k (x) over x is given by ) x k = prox γf1 (x k 1 γ f 2 (x k 1 ). Proof: We know that x k minimizes G k (x) if and only if 0 f 2 (x k ) + 1 γ (x k x k 1 ) f 2 (x k ) + f 2 (x k 1 ) + f 1 (x k ), or, equivalently, ( ) x k 1 γ f 2 (x k 1 ) x k (γf 1 )(x k ). Consequently, x k = prox γf1 (x k 1 γ f 2 (x k 1 )).

60 Proof (p.4): Theorem The sequence {x k } converges to a minimizer of the function f (x), whenever minimizers exist. Proof: G k (x) G k (x k ) = 1 2γ x x k ( f 1 (x) f 1 (x k ) 1 ) γ (x k 1 γ f 2 (x k 1 )) x k, x x k because 1 2γ x x k 2 2 g k+1(x), (x k 1 γ f 2 (x k 1 )) x k (γf 1 )(x k ).

61 Proof (p.5): Therefore, G k (x) G k (x k ) 1 2γ x x k 2 2 g k+1(x), and the iteration fits into the SUMMA class. Now let ˆx minimize f (x) over all x. Then G k (ˆx) G k (x k ) = f (ˆx) + g k (ˆx) f (x k ) g k (x k ) f (ˆx) + G k 1 (ˆx) G k 1 (x k 1 ) f (x k ) g k (x k ), so that ( ) ( ) G k 1 (ˆx) G k 1 (x k 1 ) G k (ˆx) G k (x k ) f (x k ) f (ˆx) + g k (x k ) 0.

62 Proof (p.6): Therefore, the sequence {G k (ˆx) G k (x k )} is decreasing and the sequences {g k (x k )} and {f (x k ) f (ˆx)} converge to zero. From G k (ˆx) G k (x k ) 1 2γ ˆx x k 2 2, it follows that the sequence {x k } is bounded. Therefore, we may select a subsequence {x kn } converging to some x, with {x kn 1 } converging to some x, and therefore f (x ) = f (x ) = f (ˆx). Replacing the generic ˆx with x, we find that {G k (x ) G k (x k )} is decreasing to zero. We conclude that the sequence { x x k 2 2 } converges to zero, and so {x k } converges to x. This completes the proof of the theorem.

63 Summary 1. Auxiliary-function methods can be used to impose constraints, but also to provide closed-form iterations. 2. When the SUMMA condition holds, the iterative sequence converges to the infimum of f (x) over C. 3. A wide variety of iterative methods fall into the SUMMA class. 4. The alternating minimization method is an auxiliary-function method and the 5-point property is identical to the SUMMA condition. 6. The SUMMA framework helps in proving convergence.

64 The End THE END